stowken 0.7.0

Compressed storage and retrieval of LLM token sequences
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205

//! Variable-width integer encoding for token IDs.
//!
//! Encoding scheme optimised for the Zipf distribution of token values:
//!
//! | Token value              | Bytes | First-byte pattern              |
//! |--------------------------|-------|---------------------------------|
//! | 0 – 127                  | 1     | `0xxxxxxx`                      |
//! | 128 – 16 383             | 2     | `10xxxxxx xxxxxxxx`             |
//! | 16 384 – 2 097 151       | 3     | `110xxxxx xxxxxxxx xxxxxxxx`    |
//! | 2 097 152 – 536 870 911  | 4     | `111xxxxx xxxxxxxx xxxxxxxx xxxxxxxx` |
//! | 536 870 912 – u32::MAX   | 5     | `11111111 XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX` (raw u32 BE) |
//!
//! The 5-byte escape (`0xFF` prefix) handles values beyond the 29-bit range of
//! the 4-byte scheme. All real tokenizer vocabularies fit well within the
//! 4-byte range (~512M), but the encoding must be lossless for any `u32`.

use crate::types::Token;
use thiserror::Error;

/// Errors produced during varint decoding.
#[derive(Debug, Error)]
pub enum DecodeError {
    #[error("truncated varint stream at byte {0}")]
    Truncated(usize),
    #[error("decoded token value {0} exceeds u32::MAX")]
    Overflow(u64),
}

/// Encode a slice of token IDs into a compact byte stream.
pub fn encode_tokens(tokens: &[Token]) -> Vec<u8> {
    let mut out = Vec::with_capacity(tokens.len() * 2);
    for &tok in tokens {
        encode_one(tok, &mut out);
    }
    out
}

/// Decode a byte stream produced by `encode_tokens` back to token IDs.
pub fn decode_tokens(data: &[u8]) -> Result<Vec<Token>, DecodeError> {
    let mut tokens = Vec::new();
    let mut i = 0;
    while i < data.len() {
        let (tok, consumed) = decode_one(&data[i..], i)?;
        tokens.push(tok);
        i += consumed;
    }
    Ok(tokens)
}

/// The maximum value that fits in the 4-byte encoding (29 payload bits).
const MAX_4BYTE: u32 = (1 << 29) - 1; // 536_870_911

/// Write a single token's varint encoding into `out`.
fn encode_one(tok: Token, out: &mut Vec<u8>) {
    if tok < 128 {
        // 1 byte: 0xxxxxxx
        out.push(tok as u8);
    } else if tok < 16_384 {
        // 2 bytes: 10xxxxxx xxxxxxxx
        let high = 0b1000_0000u8 | ((tok >> 8) as u8 & 0x3F);
        let low = (tok & 0xFF) as u8;
        out.push(high);
        out.push(low);
    } else if tok < 2_097_152 {
        // 3 bytes: 110xxxxx xxxxxxxx xxxxxxxx
        let high = 0b1100_0000u8 | ((tok >> 16) as u8 & 0x1F);
        let mid = ((tok >> 8) & 0xFF) as u8;
        let low = (tok & 0xFF) as u8;
        out.push(high);
        out.push(mid);
        out.push(low);
    } else if tok <= MAX_4BYTE {
        // 4 bytes: 111xxxxx xxxxxxxx xxxxxxxx xxxxxxxx  (29 payload bits)
        let high = 0b1110_0000u8 | ((tok >> 24) as u8 & 0x1F);
        let b2 = ((tok >> 16) & 0xFF) as u8;
        let b3 = ((tok >> 8) & 0xFF) as u8;
        let b4 = (tok & 0xFF) as u8;
        out.push(high);
        out.push(b2);
        out.push(b3);
        out.push(b4);
    } else {
        // 5-byte escape for values > 536_870_911:
        //   0xFF  followed by raw big-endian u32
        out.push(0xFF);
        out.extend_from_slice(&tok.to_be_bytes());
    }
}

/// Decode one varint from `data`, returning `(token, bytes_consumed)`.
fn decode_one(data: &[u8], base_offset: usize) -> Result<(Token, usize), DecodeError> {
    if data.is_empty() {
        return Err(DecodeError::Truncated(base_offset));
    }
    let first = data[0];
    if first & 0b1000_0000 == 0 {
        // 1-byte: 0xxxxxxx
        Ok((u32::from(first), 1))
    } else if first & 0b0100_0000 == 0 {
        // 2-byte: 10xxxxxx xxxxxxxx
        if data.len() < 2 {
            return Err(DecodeError::Truncated(base_offset));
        }
        let val = ((u32::from(first & 0x3F)) << 8) | u32::from(data[1]);
        Ok((val, 2))
    } else if first & 0b0010_0000 == 0 {
        // 3-byte: 110xxxxx xxxxxxxx xxxxxxxx
        if data.len() < 3 {
            return Err(DecodeError::Truncated(base_offset));
        }
        let val = ((u32::from(first & 0x1F)) << 16)
            | (u32::from(data[1]) << 8)
            | u32::from(data[2]);
        Ok((val, 3))
    } else if first == 0xFF {
        // 5-byte escape: raw big-endian u32
        if data.len() < 5 {
            return Err(DecodeError::Truncated(base_offset));
        }
        let val = u32::from_be_bytes([data[1], data[2], data[3], data[4]]);
        Ok((val, 5))
    } else {
        // 4-byte: 111xxxxx xxxxxxxx xxxxxxxx xxxxxxxx  (29 payload bits)
        if data.len() < 4 {
            return Err(DecodeError::Truncated(base_offset));
        }
        let val = ((u32::from(first & 0x1F)) << 24)
            | (u32::from(data[1]) << 16)
            | (u32::from(data[2]) << 8)
            | u32::from(data[3]);
        Ok((val, 4))
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    fn roundtrip(tokens: &[Token]) {
        let encoded = encode_tokens(tokens);
        let decoded = decode_tokens(&encoded).expect("decode failed");
        assert_eq!(tokens, decoded.as_slice(), "roundtrip mismatch for {tokens:?}");
    }

    #[test]
    fn boundary_values() {
        roundtrip(&[0]);
        roundtrip(&[1]);
        roundtrip(&[127]);
        roundtrip(&[128]);
        roundtrip(&[16_383]);
        roundtrip(&[16_384]);
        roundtrip(&[2_097_151]);
        roundtrip(&[2_097_152]);
        roundtrip(&[u32::MAX]);
    }

    #[test]
    fn empty_sequence() {
        roundtrip(&[]);
    }

    #[test]
    fn mixed_sequence() {
        let tokens: Vec<Token> = vec![0, 64, 127, 128, 1000, 16_383, 16_384, 100_000, u32::MAX];
        roundtrip(&tokens);
    }

    #[test]
    fn long_sequence() {
        let tokens: Vec<Token> = (0..10_000).map(|i| i % 150_000).collect();
        roundtrip(&tokens);
    }

    #[test]
    fn encoding_sizes() {
        // Verify expected byte widths
        assert_eq!(encode_tokens(&[0]).len(), 1);
        assert_eq!(encode_tokens(&[127]).len(), 1);
        assert_eq!(encode_tokens(&[128]).len(), 2);
        assert_eq!(encode_tokens(&[16_383]).len(), 2);
        assert_eq!(encode_tokens(&[16_384]).len(), 3);
        assert_eq!(encode_tokens(&[2_097_151]).len(), 3);
        assert_eq!(encode_tokens(&[2_097_152]).len(), 4);
        assert_eq!(encode_tokens(&[536_870_911]).len(), 4);
        // Values beyond the 4-byte range use the 5-byte escape
        assert_eq!(encode_tokens(&[536_870_912]).len(), 5);
        assert_eq!(encode_tokens(&[u32::MAX]).len(), 5);
    }

    #[test]
    fn truncated_stream_errors() {
        // A 2-byte encoding truncated to 1 byte
        let partial = vec![0b1000_0000u8]; // first byte of a 2-byte token, missing second
        assert!(matches!(decode_tokens(&partial), Err(DecodeError::Truncated(_))));
    }

    #[test]
    fn high_freq_tokens_are_small() {
        // Top tokens (0–127) should encode in 1 byte each
        let top_tokens: Vec<Token> = (0u32..128).collect();
        let encoded = encode_tokens(&top_tokens);
        assert_eq!(encoded.len(), 128, "each token 0–127 should cost 1 byte");
    }
}